A vape-detection AUC does not tell you which alerts a chosen threshold creates

Locate the score-to-alert rule

A classification system can output a score and then turn scores above a chosen boundary into alerts. Changing that boundary can change which observations receive a positive label. First identify what one observation represents and what the reference-positive label means. A time window, gesture and person are different units even when the same technology analyzes them.

Google’s original machine-learning lesson describes ROC as rates plotted across thresholds and AUC as a ranking summary. Those definitions explain the graph; they do not validate a vaping detector or establish a clinical use. A curve is not evidence that No Vape provides automatic sensing, and this guide does not recommend a threshold.

Google for Developers: Classification: ROC and AUC

Compare two fictional operating points

Imagine ten invented reference-positive windows and ninety reference-negative windows. At a stricter threshold, a toy system flags six positives and three negatives. At a more permissive threshold, it flags nine positives and eighteen negatives. More reference-positive windows are detected at the second point, while more reference-negative windows receive false flags.

The two points describe the same underlying score system under different alert rules. Neither point says what should be used in a real product. Write the selected threshold, true-positive count, missed-positive count and false-positive count on one card. Include reference-negative coverage rather than comparing raw false-alert counts across unlike datasets. All counts here are fictional arithmetic.

Ask which point the product actually uses

A paper may show an attractive curve without stating the threshold in its demonstration. An advertisement may quote AUC while users encounter a particular operating point. Look for the actual rule, whether it was chosen before the final evaluation, and whether the reported error breakdown concerns that same rule. Leave those details unresolved if the report does not provide them.

A high ranking summary does not remove the practical consequences of false entries or missed observations. The relevance depends on the stated task and handling of uncertain classifications. Ask how actual users can review an alert only if that function is documented. Do not infer a correction control or a perfectly reliable record from a favorable research metric.

Retain the curve and the operating result separately

Your final note can distinguish a reported ranking summary from the measured error pattern at the deployed threshold. If only the first is known, do not complete the second with a guessed percentage. Testing in a different setting, with different reference labels or observation units, would be another question rather than proof that every operating point transfers.

An alerting metric cannot establish reduced use, nicotine exposure or improved quitting. Those outcomes need relevant evidence of their own. For a health decision, seek qualified interpretation of the report. For ordinary technology reading, stop after identifying what was ranked, which alert rule was tested, and what practical error information remains unavailable.